[Question] Implementing background tasks via MCP servers

[NatElkins]

Hi,

Currently within a session it appears that Claude Code processes each tool call sequentially. I am wondering, is there a way to issue a long running tool call request, continue the conversation, and then accept the tool call response whenever it is ready? I was thinking that using MCP servers via SSE transport might be a way to do this, but it's not clear to me that Claude Code can handle this.

This use case is my motivation: stagewise-io/stagewise#182 (comment)

Essentially, we want to be able to inject context from an external event, without explicitly having the request/response.

Thank you!

[coygeek]

Hey, as another user, I thought I'd chime in with what I've gathered from the docs. This is a really well-thought-out question about asynchronous tool use.

From my understanding of how Claude Code currently works, the agent's tool-calling process is sequential and blocking. When Claude decides to use a tool (whether it's a built-in one like Bash or a custom one via MCP), the main conversation loop waits for that tool to complete and return a result before deciding on the next step.

So, to answer your core question:

...is there a way to issue a long running tool call request, continue the conversation, and then accept the tool call response whenever it is ready?

Currently, it seems the answer is no, not directly via MCP in the way you've described. Even with an SSE transport, the agent loop appears to wait for the entire streaming response to finish before it continues the conversation. The MCP server can't just push an unsolicited event back into the session at a later time.

A Potential Workaround (with limitations)

The good news is, there's a feature that gets part of the way there for Bash commands. The changelog for version 1.0.71 mentions:

Background commands: (Ctrl-b) to run any Bash command in the background so Claude can keep working (great for dev servers, tailing logs, etc.)

This lets you start a long-running process without blocking the main conversation.

How you might use this:

1. You could have Claude use the Bash tool to start your stagewise-mcp-server or another script as a background process. For example: ^B stagewise-mcp-server --some-task > /tmp/stagewise_output.log &
2. Claude can then continue the conversation with you immediately.
3. Later, you (or Claude) would have to explicitly check the status of that background task, perhaps by reading the output file (Read(/tmp/stagewise_output.log)).

The key limitation here is that the background task can't push its results back into the conversation. The agent has to proactively pull the information. This doesn't quite achieve the "inject context from an external event" goal you're aiming for, but it does solve the "long-running task" problem.

Regarding your architecture proposal:

Your idea of having an external process that can inject context into an active session is a classic agent architecture challenge. Based on the public docs, it doesn't seem like Claude Code's MCP implementation is designed to be a passive listener for external events. The flow is always initiated by the agent.

What you're describing sounds more like a feature request for a new capability, where Claude Code could perhaps subscribe to events from an MCP server. It's a great idea for more complex, event-driven agent workflows.

Hope this perspective helps a bit

[github-actions]

This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.

[github-actions]

This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.

[github-actions]

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.